A Framework for Efficiently Mining the OrganisationalPerspective of Business ProcessesI

Stefan Schoniga, Cristina Cabanillasb, Stefan Jablonskia, Jan Mendlingb

aUniversity of Bayreuth, GermanybVienna University of Economics and Business, Austria

Abstract

Process mining aims at discovering processes by extracting knowledge from

event logs. Such knowledge may refer to different business process perspectives.

The organisational perspective deals, among other things, with the assignment

of human resources to process activities. Information about the resources that

are involved in process activities can be mined from event logs in order to dis-

cover resource assignment conditions, which is valuable for process analysis and

redesign. Prior process mining approaches in this context present one of the

following issues: (i) they are limited to discovering a restricted set of resource

assignment conditions; (ii) they do not aim at providing efficient solutions; or

(iii) the discovered process models are difficult to read due to the number of

assignment conditions included. In this paper we address these problems and

develop an efficient and effective process mining framework that provides exten-

sive support for the discovery of patterns related to resource assignment. The

framework is validated in terms of performance and applicability.

Keywords: Business process management, declarative process mining, event

log analysis, organisational perspective, resource perspective

IThis work is funded by the “Europaischer Fonds fur regionale Entwicklung” (EFRE) undergrant 1502/89304-01/2012 (KpPQ) and the Austrian Research Promotion Agency (FFG)under grant 845638 (SHAPE).

∗Stefan SchonigEmail address: stefan.schoenig@uni-bayreuth.de (Stefan Schonig)URL: http://ai4.uni-bayreuth.de (Stefan Schonig)

Preprint submitted to Journal of Decision Support Systems June 20, 2016

1. Introduction

Business Process Management (BPM) is a well accepted method for struc-

turing the activities carried out in an organisation, analysing them for efficiency

and effectiveness, and identifying potential for improvement [1]. Processes are

not always explicitly defined when the process models are designed. Actual pro-5

cess executions may constitute a valuable input for improving process design.

Process mining provides methods for automatic process analysis, among others

for discovering processes by extracting knowledge from event logs in form of

a process model. Various algorithms are available to discover models captur-

ing the control-flow of a process, related to the behavioural perspective of the10

process [2, 3]. For perspectives like the organisational perspective, which man-

ages the involvement of human resources in processes, only partial solutions for

mining have been developed despite the importance of resource information not

only for performance but also for compliance analysis [4, 5, 6, 7].

The need to better support the organisational perspective was evidenced by15

previous approaches that mined this perspective [8, 9, 10, 11, 12, 13]. Prior

work in this area focused on discovering specific aspects of the organisational

perspective such as role models, separation of duty or social networks. However,

comprehensive and integrated support for the well-established workflow resource

patterns, and specifically in this context for the so-called creation patterns [14],20

was missing. Furthermore, the close interplay between the organisational and

the behavioural perspectives was disregarded [15]. In [16] we addressed these

gaps by developing a declarative process mining approach for the organisational

perspective, which supports all the creation patterns as well as what we called

cross-organisational patterns, which discover how the involvement of resources25

influences the control-flow of the process.

The research reported in this paper extends our prior work towards an effi-

cient and effective mining framework. As illustrated in Figure 1, the framework

is divided into an event log pre-processing phase, a phase for integrated resource

mining including cross-perspective patterns, and a model post-processing phase.30

2

Discovery of resource assignment

rules and influence on control flow

Org. Model

Improving mining efficiency by

generating only reasonable

candidates

Improving understandability of

results by pruning redundant

rules

Event Log

Pre-Processing

Model

Post-Processing

Log

Integrated Rule-based

Resource Mining

Log

Workflow

Resource Patterns

Cross-perspective

Patterns

Figure 1: Framework for discovering resource-aware, declarative process models

We evaluate our approach with an implementation of the three phases; with sim-

ulation experiments for measuring performance; and with the application of the

approach on a real-life event log for checking its effectiveness.

This research extends our previous work [16] as follows: (i) the developed

pre-processing method increases the efficiency of the approach; (ii) the devel-35

oped post-processing techniques increase the understandability of the results;

(iii) a prototype of the entire framework has been implemented using Drools;

and (iv) the approach has been extensively validated. In addition, the mining

approach is explained in more detail. With our work, we complement research

on process mining with an extensive support of the organisational perspective.40

The remainder of this paper is structured as follows: Section 2 introduces

background information. Section 3 describes our process mining approach. Sec-

tions 4 and 5 describe the event log preprocessing and postprocessing phases

of the framework, respectively. Section 6 explains the evaluations performed.

Section 7 describes the related work and Section 8 concludes the paper.45

2. Background

In the following we introduce the concepts upon which our approach has

been developed.

3

2.1. Organisational and Cross-Perspective Patterns in Processes

The well-known workflow resource patterns [14] capture the various ways in50

which resources are represented and utilised in business processes. Of specific

interest to our research are the creation patterns since they describe different

ways in which resources can be assigned to activities. These patterns, which

will be referred to as organisational patterns from now on, include: Direct Dis-

tribution, or the ability to specify at design time the identity of the resource55

that will execute a task. Role-Based Distribution, or the ability to specify at

design time that a task can only be executed by resources that have a given role.

Organisational Distribution, or the ability to offer or allocate activity instances

to resources based on their organisational position and their organisational rela-

tionship with other resources. Separation of Duties, or the ability to specify that60

two tasks must be allocated to different resources in a given process instance.

Case Handling, or the ability to allocate all the activity instances within a

given process instance to the same resource. Retain Familiar (a.k.a. Binding

of Duties), or the ability to allocate an activity instance within a given pro-

cess instance to the same resource that performed a preceding activity instance.65

Capability-Based Distribution, or the ability to offer or allocate instances of an

activity to resources based on their specific capabilities. Deferred Distribution,

or the ability to defer the specification of the identity of the resource that will

execute a task until run time. History-Based Distribution, or the ability to offer

or allocate activity instances to resources based on their execution history. Note70

that the creation patterns Authorisation and Automatic Execution are not in

the list because they are not directly related to resource assignment.

It has been identified that process control-flow is intertwined with depen-

dencies upon resource characteristics [15]. For instance, sometimes an activity

must be executed eventually before another one for specific resources but not for75

others. As an example, resources with a certain role (e.g., trainees) must always

perform a certain activity (e.g., double-check result) before they can continue

with the following activity, but this might not be required for other roles (e.g.,

supervisors). We call this pattern Role-Based Sequence. A specific collection of

4

such cross-perspective patterns capturing these situations has not been defined.80

Nonetheless, in general, they can be defined by combining the aforementioned

organisational patterns with the control-flow patterns described in [18]. The

Resource-Based Response pattern, e.g., describes that for a special resource a

certain activity has to follow eventually on another activity.

The organisational and the cross-perspective patterns constitute the set of85

patterns to be discovered by our framework.1

2.2. Event Logs for Mining the Organisational Perspective

Our mining approach takes as input (i) an event log, i.e., a machine-recorded

file that reports on the execution of tasks during the enactment of the instances

of a given process; and (ii) organisational background knowledge, i.e., prior90

knowledge about the roles, capabilities and the membership of resources to

organisational units, among others. In an event log, every process instance cor-

responds to a sequence (trace) of recorded entries, namely, events. We require

that events contain an explicit reference to the enacted task and to the operating

resource. Both conditions are commonly respected in real-world event logs [2].95

For instance, the following excerpt of a business trip process event log encoded

in the XES logging format [17] shows the recorded information of the start event

of activity Apply for trip performed by resource ST.

<event>

<string key="org:resource" value="ST"/>100

<date key="time:timestamp" value="2013-08-06T14:58:00.000+01:00"/>

<string key="concept:name" value="Apply for trip"/>

<string key="lifecycle:transition" value="start"/>

</event>

2.3. Representing the Output of the Mining105

Since our aim is to discover the patterns explained in Section 2.1, the mod-

elling language to represent the discovered processes must offer the possibility

1Therefore, when we speak about mining the organisational perspective we refer to both

sets of patterns.

5

hasRolehasRole

Superv isor

Superv isor

Professor Student

ST BRSJ ...

predicate

object

Relation Entity

Identity GroupRelationType

subject

(a) Organisational meta model

hasRolehasRole

Superv isor

Superv isor

Professor Student

ST BRSJ ...

predicate

object

Relation Entity

Identity GroupRelationType

subject

(b) Organisational model

Figure 2: Organisational meta model and example organisational model

to define (i) expressive organisational patterns and (ii) cross-perspective pat-

terns. Two different representational paradigms for process models can be dis-

tinguished: procedural models describe which activities can be executed next110

in a process, and declarative models define by means of rules the execution con-

straints that the process has to satisfy [18]. Current procedural languages like

Business Process Model and Notation (BPMN) [19] put a strong emphasis on

control-flow and assume other perspectives to be specified separately. Cross-

perspective patterns cannot be readily modelled. Declarative process modelling115

does not limit the number of perspectives involved in the constraints defined.

However, a central shortcoming of existing languages like Declare [18] is that

they are not provided with the capability to directly define the connection be-

tween the process behavior and other perspectives. We will use the Declarative

Process Intermediate Language (DPIL) [20] for modelling the output of the min-120

ing because it supports multiple perspectives including the behavioural and the

organisational perspectives, as well as the interplay between them. DPIL is ex-

pressive enough to cover the workflow patterns [20]. Nonetheless, the concepts

of our approach are generic such that other declarative languages, such as Sciff

[21] or LTL-based formalisms [22], could also be used as long as they provided125

support for the modelling of our target patterns.

In order to express organisational information, DPIL builds upon a generic

organisational meta model [23] that is depicted in Figure 2a. It comprises the

following elements: Identity represents agents that can be directly assigned

6

to activities, i.e., both human and non-human resources. Group represents130

abstract agents that may describe several identities as a whole, e.g., roles or

groups. Relation represents the different relations (RelationType) that may

exist between these elements. It is well suited for defining, e.g., that an identity

has a specific role, that a person is the boss of another person, or that a person

belongs to a certain department. In this context, relations are generally irreflex-135

ive. A relation is irreflexive if an identity cannot be in relation to itself. The

supervisor relation, e.g., is irreflexive, since a person cannot be their own super-

visor. In addition, some relations may be transitive. A relation is transitive if

whenever an individual i1 is related to another individual i2 with that relation,

and i2 is in turn related to a third individual i3 with the same relation, then140

i1 is also related to i3. For instance, the supervisor and delegate relations are

typically transitive because organisations are usually hierarchically structured.

Figure 2b illustrates an exemplary organisational model of a university research

group, composed of two roles (Professor, Student) assigned to three people (SJ,

ST, BR) and two relations between them indicating who is supervised by whom.145

DPIL provides a textual notation based on the use of macros to define

reusable rules. For instance, the sequence(a,b) macro states that the existence

of a start event of task b implies the previous occurrence of a complete event of

task a; and the role(a,r) macro states that an activity a is assigned to a role

r. Figure 3 shows an example of a process for trip management modelled with150

DPIL. It specifies that it is mandatory to approve a business trip before flight

tickets can be booked. Moreover, it is necessary that the approval be carried

out by a resource with the role Professor.

3. Mining the Organisational Perspective

In this section we describe our approach to discover organisational and cross-155

perspective patterns. First, we describe how rule candidates are generated and

checked. Then, we classify them according to support, confidence and interest

factor values. Finally, we present a catalogue of rule templates that covers the

7

use group Professor

process BusinessTrip {

task Book flight

task Approve Application

ensure role(Approve Application, Professor)

ensure sequence(Approve Application, Book flight)

}

Figure 3: Process for trip management modelled with DPIL

target expressiveness (cf. Section 2.1).

3.1. Generation and Checking of Rule Candidates160

Declarative process modelling languages like DPIL are based on so-called

rule templates. A rule template captures frequently needed relations and defines

a particular type of rules. Templates have formal semantics specified through

logical formulae and are equipped either with user-friendly graphical represen-

tations (e.g., in Declare) or macros in textual languages (e.g., in DPIL). Unlike165

concrete rules, a rule template consists of placeholders, i.e., typed variables. A

rule template is instantiated by providing concrete values for these placeholders.

For instance, the model described in Section 2 makes use of two rule templates

represented by the macros sequence(T1,T2) and role(T ,G). These templates

comprise placeholders of type Task T as well as Group G. In all well-known170

declarative process mining approaches, rule templates are used for querying the

provided event log to find solutions for the placeholders. A solution is any com-

bination of concrete values for the placeholders that yields a concrete rule that

is satisfied in the event log. First, all possible rules need to be constructed by

instantiating the given set of rule templates with all possible combinations of175

occurring process elements provided in the event log. For example, the sequence

template consists of two placeholders of type Task. Assuming that |T | different

tasks occur in the event log, |T |2 rule candidates are generated.

Let |Θ| be the number of different rule templates to be checked and |Pj(i)|

the number of different elements in the event log of a certain parameter type

Pj(i) contained in rule template θi. Let k(i) be the number of placeholders in θi.

8

Trace start(of t1) direct(t1,i1)

{s(t2,i1), c(t2,i2), s(t3,i1), c(t3,i1)} false false

{s(t1,i1), c(t1,i1), s(t2,i2), c(t2,i2), s(t3,i1), c(t3,i1)} true true

{s(t1,i1), c(t1,i1), s(t3,i3), c(t3,i3), s(t2,i2), c(t2,i2)} true true

{s(t1,i1), c(t1,i1), s(t3,i3), c(t3,i3), s(t2,i2), c(t2,i2)} true true

{s(t1,i4), c(t1,i4), s(t3,i1), c(t3,i1)} true false

Table 1: Event log and satisfaction of an example rule and its condition

The number of generated rule candidates |RCand| is |P1(1)| · |P2(1)| · ... · |Pk(1)(1)|

+ |P1(2)| · |P2(2)| · ... · |Pk(2)(2)| + ... + |P1(i)| · |P2(i)| · ... · |Pk(i)(i)| and therefore,

|RCand| =|Θ|∑i=1

(

k(i)∏j=1

|Pj(i)|) (1)

The resulting candidates are subsequently checked w.r.t. the log. In many

cases a rule candidate can be trivially valid. Consider the candidate direct(t1,i1),180

i.e., start(of t1) implies start(of t1 by i1), which holds when task t1 is performed

by identity i1, and the event log shown in Table 1. The notation used encodes

the start and complete events of a specific task t performed by an identity i

with s(t,i) and c(t,i), respectively. The given events are ordered temporally so

that timestamps are not encoded explicitly. In the first trace the rule holds185

trivially because t1 never happens. Using the terminology of [24], we say that

the rule is vacuously satisfied. It is necessary to discriminate between traces

in which a rule is trivially true and traces in which the rule is non-vacuously

satisfied. Only the latter are considered interesting [25]. For first order logic

rules that depict implications of the form A → B, trivially and non-vacuously190

valid rules can be discriminated by additionally checking the condition A of

the rule separately. Table 1 shows the results of checking the non-vacuous

satisfaction of the direct(t1,i1) rule as well as its condition for each trace of the

example log. In the first trace the rule is not (non-vacuously) satisfied because

t1 is never started, i.e., the condition is false. The rule holds non-vacously in195

the traces two to four. It is violated in trace five.

9

3.2. Metrics to Classify Rule Candidates

Checking rule candidates as described above provides for every candidate the

number of instances, i.e., the traces in the event log where it non-vacously holds.

Based on these values it is possible to classify rules and to separate non-valid

from valid ones. Maggi et al. [24] adopted different metrics, specifically support

(supp), confidence (conf ) and interest factor (int) proposed by association rule

mining for evaluating the relevance of rule candidates. Let |Φ| be number of

traces in an event log Φ. Let |σnv(r)| be the number of traces in which a rule

r : A → B is non-vacously satisfied. The support supp(r), confidence conf(r)

and int(r) values of a rule r are defined as:

supp(r) :=|σnv(r)||Φ|

, conf(r) :=supp(r)

supp(A), int(r) :=

supp(r)

supp(A) · supp(B)(2)

Considering again the event log of Table 1 and the direct(t1,i1) rule. Its

support evaluates to supp(r) = 0.6, its confidence to conf(r) = 0.75 and its

interest factor to int(r) = 1.25. We make use of the confidence value to classify200

a rule candidate r as a valid rule (i.e., satisfied in almost all traces) or a non-

valid rule (i.e., violated in most of the recorded traces). Therefore, the threshold

minConf is introduced to classify rule candidates. Candidates r with conf(r) >

minConf are classified as valid. All rule candidates r with conf(r) < minConf

are non-valid rules and are not part of the resulting process model. Note that in205

case of rules that do not depict implications, the condition is satisfied in every

trace; therefore, supp(A) = 1 and conf(r) = supp(r). Using the confidence

values of rule candidates it is directly possible to generate a DPIL process model

reflecting organisational and cross-perspecitve patterns.

3.3. Rule Templates for Mining the Organisational Perspective210

Since DPIL builds upon a flexible organisational meta model (cf. Sec-

tion 2.3), it is possible to define rule templates that describe many aspects of the

organisation. By instantiating these rule templates with all possible parameter

combinations of defined resources, groups and relation types, it is possible to

10

generate rule candidates that focus on the organisational perspective of the pro-215

cess to be analysed. These candidates can then be checked under consideration

of the event log and the organisational model.

In the following we define rule templates and their macros for our target set

of patterns. First of all, we distinguish between templates for organisational

patterns and templates for cross-perspective patterns. The former are, in turn,220

divided into two groups based on the types and number of parameters: rule

templates related to a single task and rule templates related to more than one

task. We provide representative examples for each group of rule templates

that cover frequently needed organisational information. Note that besides the

templates described next, further templates could be defined individually to225

cover the analyst’s needs.

3.3.1. Rule Templates for the Assignment of Resources to a Single Task

This group includes rule templates that define organisational patterns re-

ferred to one process activity. The Direct Distribution pattern can be extracted

with a direct(T ,I) template. Given the free variables T and I and an event log230

with |T | distinct tasks and |I| distinct resources, there are |T | · |I| candidates to

be checked.

direct(T,I) iff start(of T) implies start(of T by I)

The Role-Based distribution pattern can be extracted with a role(T ,G) tem-

plate. Here, rule candidates for every task and group combination are generated,235

i.e., |T | · |G| rule candidates need to be checked.

role(T,G) iff start(of T by :p) implies

relation(subject p predicate hasRole object G)

The Capability-Based distribution pattern can be extracted with a capability(T ,RT ,G)

template. A capability is represented by a relation of an individual to a group,240

e.g., i1 hasDegree ComputerScience. According to the placeholders, |T |·|RT |·|G|

candidates are generated.

11

capability(T, RT, G) iff

start(of T by :p) implies relation(subject p predicate RT object G)

The assignment of resources based on organisational positions of individuals,245

described by the Organisation-Based Distribution pattern, can be extracted with

an orgDistSingle(T ,RT ,G) template. Here, |T | · |RT | · |G| rules must be checked.

orgDistSingle(T, RT, G) iff

start(of T by :p) implies relation(subject p predicate RT object G)

3.3.2. Rule Templates for the Assignment of Resources to Several Tasks250

This group includes rule templates that define organisational patterns re-

ferred to several tasks. The Separation of Duties pattern can be extracted

with a separate(T1,T2) template. For this template, |T |2 candidates need to be

checked.

separate(T1,T2) iff start(of T1 by :p) and start(of T2) implies255

start(of T2 by not p)

The Retain Familiar pattern can be extracted with a binding(T1,T2) tem-

plate. Similarly to the previous case, |T |2 candidates need to be checked.

binding(T1,T2) iff start(of T1 by :p) and start(of T2) implies

start(of T2 by p)260

The Case Handling pattern can be extracted with a caseHandling template.

Here, |T | candidates have to be checked.

caseHandling iff forall(task T start(of T) implies start(of T by :p))

Resources can also be assigned to tasks according to their organisational

relation with the performers of other process activities, e.g., an approval task265

might be assigned to people that can supervise the work done by the performers

of a previous task. This is covered by the Organisation-Based Distribution

pattern and can be extracted with an orgDistMulti(T1,T2,RT ) template where

variable RT specifies the type of relation between the two individuals involved.

There exist |T |2 · |RT | rule candidates.270

12

orgDistMulti(T1,T2,RT) iff start(of T1 by :p1) and start(of T2 by :p2)

implies relation(subject p1 predicate RT object p2)

3.3.3. Cross-Perspective Rule Templates

A cross-perspective rule describes a temporal dependency or constraint be-

tween tasks but only applies for a certain set of identities, like in the following275

examples. Note, that other well-known control-flow patterns described in [18]

can be defined in a similar way. The Role-Based Sequence pattern can be ex-

tracted with a roleSequence(T1,T2,G) template. Here, |T |2 · |G| candidates need

to be checked.

roleSequence(T1,T2,G) iff start(of T2 by :p at :t) and280

relation(subject p predicate hasRole object G)

implies complete(of T1 at < t)

The Resource-Based Response pattern can be extracted with a resourceResponse(T1,T2,I)

template. In this case, |T |2 · |I| candidates need to be checked.

resourceResponse(T1,T2,I) iff complete(of T1 by I at :t) implies285

start(of T1 at > t)

4. Pre-processing to Extract Meaningful Parameters

Real-life event logs and organisational models potentially contain a big set of

distinct tasks, resources and groups. For instance, the BPI challenge 2011 event

log of a hospital information system [26] contains 623 different tasks and 42290

organisational groups. By only considering the role template, this already leads

to 623·42 = 26166 candidates to be checked. Although many of these parameter

combinations never occur together in the same trace, the corresponding rules

need to be checked. This problem can also be observed when considering task-

resource combinations of the event log in Table 1. Resource i4 only occurs295

together with task t1. Hence, candidates of the direct template where I = i4

and T 6= t1 are trivially true in all traces and can be neglected without checking.

The method proposed in [24] uses the well-known Apriori algorithm to pre-

process the log and to extract task combinations that frequently occur together.

13

Rule Template Item Itemset

direct(T, I) (Task, Identity) L1: {(Task, Identity)}

role(T, G) (Task, Group) L1: {(Task, Group)}

capability(T, RT, G) (Task, Group) L1: {(Task), (Group)}

orgDistS(T, RT, G) (Task, Group) L1: {(Task), (Group)}

binding(T, T) (Task) L2: {(Task), (Task)}

separate(T, T) (Task) L2: {(Task), (Task)}

orgDistMulti(T, T, RT) (Task, Task) L2: {(Task), (Task)}

roleSequence(T, T, G) (Task, Group) L2: {(Task, Group), (Task, Group)}

Table 2: Required itemsets for exemplary organisational rule templates

The problem of mining frequent itemsets is to find all itemsets that satisfy a

user-specified minimum support. The support of an itemset X is the percentage

of traces that contain the items of X. Note that this support value is different

from the one defined in Section 3.2, which depicts the fraction of traces where a

certain rule is non-vacuously satisfied. Specifically, let |Φ| be the total number

of traces recorded in the log. Let σX be the set of traces that contain a set of

items X. The support value of an itemset X in Φ is defined as

supp(X) =|σX ||Φ|

, where σX = {σ ∈ Φ|∀x∈X x ∈ σ} (3)

A task combination is considered to be relevant if it occurs in a sufficient

number of traces, i.e., if its support value is greater than a given threshold

minSupp. A minSupp of 0.05, e.g., claims that only rule candidates whose pa-300

rameter combinations occur in at least 5% of the recorded traces are considered.

We extended this method to also extract task-resource and task-group combi-

nations that frequently occur together. In this way, it is possible to reduce the

number of organisational rule candidates by ignoring infrequent parameter com-

binations. For instance, for the example log, only one out of three direct(T ,i4)305

candidates is generated and checked.

Table 2 shows the form of a single item and the required itemset for the

already defined rule templates (cf. Section 3.3). Regarding the rule templates

14

for the assignment of resources to a single task, since only one task is involved,

itemsets X with |X| = 1 are required. For instance, the direct template has two310

placeholders, one for tasks and one for identities and hence, itemsets of the form

(Task, Identity) are needed. Regarding the rule templates for the assignment

of resources to several tasks, since in all these templates two tasks are involved,

itemsets with |X| = 2 are required. The binding template, e.g., takes frequent

items of the form (Task, Task). Finally, the cross-perspective rule templates315

also have two placeholders and hence, itemsets with |X| = 2 are required. The

templates capability, orgDistS and orgDistM additionally contain a variable for

a relation type. The amount of different relation types in organisational models,

however, is usually insignificant compared to the number of different individuals

and groups and can therefore be neglected.320

5. Pruning of Discovered Models

The output of the mining phase is a process model with rules that state which

resources are assigned to the process tasks, e.g., resources with specific roles or

capabilities. The mining method extracts all the assignment rules related to

each task. However, when several rules are extracted for one single task, not325

all of them might be strictly necessary to understand the process. Specifically,

some rules may be implied by stronger rules because they are less restrictive

and do not provide any value to the current resource assignment expression of

a task. Those rules complicate the understandability of discovered models and

hence, they are unnecessary. We identified two pruning approaches to eliminate330

unnecessary rules: pruning based on organisational rule hierarchies and pruning

based on transitive reduction. The requirement for all pruning operations is that

they do not change the meaning of the generated model.

5.1. Pruning based on Organisational Rule Hierarchies

Maggi et al. [27] proposed a technique to post-process a discovered model335

and to remove weaker rules if they are already implied by stronger rules only

15

focusing on the hierarchy of control-flow templates. Hierarchies also exist in

case of organisational rules.

We define rule hierarchies for the rule templates defined in Section 3.3. For

that purpose, we introduce the dominates relation →dom between two rules r1340

and r2. Specifically, r1 →dom r2 means that rule r1 is stronger than rule r2.

The defined rule hierarchies can then be used to prune and simplify discovered

models. If a model contains two assignment rules r1 and r2 concerning the

same task and r1 →dom r2, then r2 can be pruned, i.e., removed from the

model. User-defined rule types have to be integrated in exiting hierarchies by345

modelling experts. In order to justify the rule hierarchies described next, the

following sets and functions must be introduced: T = {t1, t2, ..., tn} is a set

of tasks; Ri = {r1, r2, ..., rm} is a set of assignment rules discovered for task

ti; I = {i1, i2, ..., ip} is a set of identities (i.e., individuals) of an organisation;

G = {g1, g2, ..., gq} is a set of user groups of an organisation (e.g., roles); id :350

Ri → I returns the set of identities that meet the conditions defined by a rule;

pp : T → I returns the set of potential performers of a task, where pp(ti) =⋂Ri

and pp(ti) 6= ∅ because otherwise rules would not have been extracted from the

event log for task ti; and ap : T → I returns the actual performer of a task for

a specific task instance, so that ap(ti) ∈ pp(ti).355

We next explain how the rule hierarchies have been derived, providing a

demonstration and an example for each dominates relation identified.

5.1.1. Rule Hierarchy for the Templates referred to a Single Task

We first focus on resource assignment rules for a single task (cf. Sec-

tion 3.3.1), represented as Θ1 = {direct(T ,I), role(T ,G), capability(T ,RT ,G),360

orgDistSingle(T ,RT ,G)}. Next, we describe and demonstrate the domination

relations found out in Θ1. For that, let us imagine that we have discovered two

rules R1 = {r1, r2} for task t1. Therefore, pp(t1) = id(r1) ∩ id(r2), pp(t1) 6= ∅.

The aim in all cases is to prove that id(r1) ⊆ id(r2), i.e., the individuals of r1

are a subset of the individuals of r2 and hence, r2 is weaker and can be removed.365

The resulting rule hierarchy is visualised in Fig. 4a.

16

CASE HANDLING

do

m

DIRECT ALLOCATION

ROLE BASED

ALLOCATION

CAPABILITY BASED

ALLOCATION

ORGANISATIONAL

ALLOCATION

domdom

ROLE(T,G) CAPABILITY(T,RT,G) ORGDISTS(T,RT,G)

DIRECT(T,I)

CASEHANDLING

ORGANISATIONAL

ALLOCATION

do

m

SEPARATION OF

DUTIES

CASE HANDLING

BINDING OF DUTIES

domdom

CASEHANDLING

BINDING(T1,T2)

SEPARATE(T1,T2)

ORGDISTM

(T1,T2,RT)

(a) Assignment rules w.r.t. a single task

CASE HANDLING

do

mDIRECT ALLOCATION

ROLE BASED

ALLOCATION

CAPABILITY BASED

ALLOCATION

ORGANISATIONAL

ALLOCATION

domdom

ROLE(T,G) CAPABILITY(T,RT,G) ORGDISTS(T,RT,G)

DIRECT(T,I)

CASEHANDLING

ORGANISATIONAL

ALLOCATION

do

m

SEPARATION OF

DUTIES

CASE HANDLING

BINDING OF DUTIES

domdom

CASEHANDLING

BINDING(T1,T2)

SEPARATE(T1,T2)

ORGDISTM

(T1,T2,RT)

(b) Assignment rules w.r.t. two tasks

Figure 4: Hierarchies of organisational patterns

direct(t1,i1) →dom role(t1,g1), direct(t1,i1) →dom capability(t1,rt1,g1),

direct(t1,i1) →dom orgDistS(t1,rt1,g1). Direct rules dominate role rules, ca-

pability rules and orgDistS rules. The demonstration of the three relations

is the same, being r1 = direct(t1,i1) in all cases and r2 = role(t1,g1), r2 =370

capability(t1,rt1,g1) and r2 = orgDistS(t1,rt1,g1), respectively.

Proof. We demonstrate that id(r1) ⊆ id(r2) by contradiction. Let id(i1) =

{r1} and id(r2) = {i2, i3, i4}, so id(r1) 6⊆ id(r2). That means i1 does not have

role g1. Then, pp(t1) = id(r1)∩id(r2) = ∅, which is not possible by definition, as

aforementioned. Therefore, and since |id(r1)| = 1, id(r1) ⊆ id(r2) is mandatory375

and hence, pp(t1) = id(r1), which means r2 is redundant and can be removed.

Example. Consider that a specific task Book flight has always been per-

formed by a resource ST who has the role Student according to the organisa-

tional model. Then, the proposed method will (inevitably) discover rules di-

rect(Book flight,ST) and role(Book flight,Student). The identities derived from380

the latter rule are ST and BR. However, there is no evidence that BR can

execute the task and hence, the role rule is not strong enough to be considered

in the resource assignment.

role(t1,i1) 6↔dom capability(t1,rt1,g1), role(t1,i1) 6↔dom orgDistS(t1,rt1,g1),

capability(t1,rt1,g1) 6↔dom orgDistS(t1,rt1,g1). There is no domination re-385

lation between role and capability rules, role and orgDist rules, and capability

17

and orgDist rules. The demonstration is equivalent for any r1 and r2 belonging

to these three groups.

Proof. The difference with respect to the previous demonstration lies on

the cardinality of the rules involved. In this case, for any r1, r2 of one pair390

of rule types, |id(r1)| >= 1 and |id(r2)| >= 1. Since id(r1) ∩ id(r2) 6= ∅, then

either id(r1) ⊆ id(r2) or id(r2) ⊆ id(r1) depending on the number of individuals

meeting the conditions specified by the rules. Therefore, a subsumption rela-

tion cannot be generalised and hence, both rules are, in general, necessary to

calculate the potential performers of a task t1, such that pp(t1) = id(r1)∩id(r2).395

Example. Consider the situation where the rules role(Approve application,Professor)

and capability(Approve application,hasDegree,CS) have been extracted. It means

that the task has been performed by someone with the role Professor and with

a degree in Computer Science (CS). However, there might also be professors

that do not have a degree in Computer Science, and vice versa. Therefore, to400

describe the necessary task condition, both rules are needed.

5.1.2. Rule Hierarchy for the Templates referred to Several Tasks

We now focus on resource assignment rules that involve two different tasks

(cf. Section 3.3.2), represented as Θ2 = {binding(T1,T2), separate(T1,T2), orgDistMulti(T1,T2,RT )}.

Next, we describe and demonstrate the domination relations found out in Θ2.405

For that, let us imagine that we have discovered two rules R1 = {r1, r2} for

task t1, where one of the rules, in turn, refers to the assignment rule of task t2.

Similarly to the previous case, pp(t1) = id(r1) ∩ id(r2), pp(t1) 6= ∅. The aim is

again to prove that id(r1) ⊆ id(r2), i.e., the individuals of r1 are a subset of the

individuals of r2 and hence, r2 is weaker and can be removed. The resulting410

rule hierarchy is visualised in Fig. 4b.

separate(t1,t2) 6↔dom binding(t1,t2). There is no domination relation be-

tween separate and binding rules.

Proof. The demonstration is a contradiction by definition. The separate

rule implies that ∀ap(t1),∀ap(t2) in a specific process instance, ap(t1) 6= ap(t2),415

18

i.e., both tasks have always been performed by different identities. The binding

rule, however, states that ∀ap(t1),∀ap(t2) in a specific process instance, ap(t1) =

ap(t2), i.e., both tasks have always been performed by the same identity. In case

both rules were extracted for task t1, id(r1) ∩ id(r2) = ∅ and hence, pp(t1) = ∅.

Therefore, these two rules can simply never be extracted at the same time420

because they are mutually exclusive.

orgDistMulti(t1,t2,rt1) 6↔dom binding(t1,t2). There is no domination rela-

tion between orgDistMulti2 and binding rules.

Proof. Similarly to the previous case, the demonstration is a contradiction

by definition. With an orgDistMulti rule using an irreflexible relation, ap(t1) 6=425

ap(t2). However, according to the binding rule, ap(t1) = ap(t2). Hence, rules

of these two types will never be extracted at the same time because they are

mutually exclusive.

Example. Consider the situation where the rules orgDistMulti(Approve ap-

plication,Apply for trip,supervisor) and binding(Approve application,Apply for430

trip) have been extracted for a task. It means that the application must be ap-

proved by the supervisor of the person who applies for the trip. Since a person

cannot be a supervisor of herself, the tasks are performed by different individu-

als. However, according to the second rule, the two tasks should be performed

by the same person.435

orgDistMulti(t1,t2,rt1)→dom separate(t1,t2). orgDistMulti rules dominate

separate rules.

Proof. Let r1 = orgDistMulti(t1,t2,rt1) and r2 = separate(t1,t2). Assum-

ing irreflexible relations in the organisation, according to both rules ap(t1) 6=

ap(t2). Since id(r1) ⊆ id(r2), pp(t1) = id(r1), which means r2 is redundant and440

can be removed.

Example. Consider the situation where the rules orgDistMulti(Approve ap-

plication,Apply for trip,supervisor) and separate(Approve application,Apply for

2Note that we assume that all relations are irreflexive (cf. Section 2).

19

trip) have been extracted for a task. It means that the application must be ap-

proved by the supervisor of the person who applies for the trip. Since a person445

cannot be a supervisor of herself, the tasks are performed by different individu-

als. However, not all the other persons in the organisation might be supervisors

of the person applying for the trip. Therefore, this condition is more restrictive

than the separation of duties and then, the latter is not necessary in the resource

assignment expression.450

5.1.3. Rule Hierarchy for the Cross-Perspective Templates

Finally, we address cross-perspective rules (cf. Section 3.3.3), represented

as Θ3 = {roleSequence(T1,T2,G), resourceSequence(T1,T2,I)}. Notice that in

this case the approach is different from Θ1 and Θ2 since we aim at generalising

under which conditions a specific activity order must take place. That means455

that a rule r1 is stronger than a rule r2 if id(r2) ⊆ id(r1). As demonstrated

next, roleSequence(t1,t2,g1) →dom resourceSequence(t1,t2,i1).

Proof. Let us imagine that we have discovered two rules R1 = {r1, r2},

where r1 = roleSequence(t1,t2,g1)3 and r2 = resourceSequence(t1,t2,i1). The

temporal dependency is the same in both cases, specifically, a specific task460

order determined by sequence(t1,t2). Therefore, we could assume that r1 =

role(t1,g1) and r2 = direct(t1,i1). According to the aforementioned criterion,

since |id(r1)| >= 1 and |id(r2)| = 1, id(r2) ⊆ id(r1), i.e., the individuals of r2

are a subset of the individuals of r1 and hence, r2 is weaker and can be removed.

Example. Consider that task Apply for trip has always been performed465

before task Book flight when executed either by resource ST or by resource

BR, who have the role Student according to the organisational model. Then,

the proposed method will (inevitably) discover rules resourceSequence(Apply

for trip,Book flight,ST), resourceSequence(Apply for trip,Book flight,BR) and

roleSequence(Apply for trip,Book flight,Student). Since the individuals of both470

3Note that to discover a roleSequence it is necessary to identify at least two entries in the

log in which different resources with the same role are associated to a specific task sequence.

20

resourceSequence rules (i.e., ST and BR) are a subset of the individuals of the

roleSequence rule, they can both be removed from the model.

5.2. Pruning based on Transitive Reduction

The assignment rules in Θ2 (cf. Section 5.1.2) may be affected by transitivity.

In particular, redundancy may be caused by the interplay of three or more rules475

of the same type applied to different activities. Consider a set of discovered

binding rules, such as binding(t1,t2), binding(t2,t3) and binding(t1,t3). Here,

the rule between t1 and t3 is redundant because it belongs to the transitive

closure of the other rules. In other words, if task t1 has always been performed

by the same resource as t2, and task t3 has always been performed by the480

same resource as t2, then also t1 and t3 have been performed by the same

resource. Therefore, binding(t1,t3) is unnecessary and could be removed using

the transitive reduction algorithm as defined in [28]. OrgDistMulti rules can be

transitively reduced in a similar way if they refer to the same relation type rt

and if rt is a transitive relation (cf. Section 2). However, separate rules are not485

transitive, i.e., if t1 is not performed by the same resource as t2 and t2 is not

executed by the same resource as t3, then we cannot conclude that t1 is also not

performed by the same resource as t3.

6. Evaluation

We evaluate our framework in three steps. We first describe how it has been490

implemented. We then show its efficiency with simulation experiments. Finally,

we report on the results of applying the framework on a real-life event log.

6.1. Implementation

The problem of checking a large set of rule candidates can be solved by

efficient pattern matching methods like the rete algorithm [29]. Instead of495

checking each rule separately, the rete algorithm first identifies common parts

of the provided set of rules and constructs a rete network. Based on this

decision network, common rule parts just need to be checked once. The JBoss

21

Drools platform4 provides a current implementation of this method. In order

to check rule candidates with Drools, they are translated into the Drools Rule500

Language (DRL). Like in DPIL, rules in DRL consist of a condition (when

part) and a consequence (then part). If the condition holds, the consequence

will be performed. DRL supports language elements to describe rules of first

order logic, hence being equivalent to DPIL. The transformation of the most

important expressions from DPIL to DRL are shown in Table 3. DPIL rules505

are translated into DRL rules like in row 3. As can be seen, the complete DPIL

rule is placed in the when part of the DRL rule. The consequence, i.e., the

then part, only contains a procedure call that signals the satisfaction of the

corresponding rule to the program environment (listener). Since DRL does not

support a logical implication directly, DPIL implications must be translated510

into DRL according to the logical equivalence A → B ≡ ¬(A ∧ ¬B) (cf. row 4

in Table 3). The described approach has been implemented in the DpilMiner

application5.

6.2. Performance Evaluation

To analyse performance we used the DpilMiner with different configurations515

using an event log of a university business trip management system6. The

log contains 2104 events of 10 different activities related to the application

4Documentations about JBoss Drools is available at http://docs.jboss.org/drools5A screencast of the DpilMiner is accessible at http://www.kppq.de/miner.html6The event log is available for download at http://workbench.kppq.de

Nr. DPIL expression DRL expression

1 task T :t $t: Task(id == “T”)

2 start(of T) $t: Task(id == “T”) and Start(Task == $t)

3 expr rule Id

when expr then listener.onRuleOccured(drools.getRule()));

4 x implies y not (x and not y)

Table 3: Rules for transforming DPIL to DRL expressions

22

300

243241 229

222

22 22 22 20 1914 14 14 12 11

7.75

6.44 6.446.12

5.36.74

5.445.23 5.08 4.9

0

1

2

3

4

5

6

7

8

9

0

50

100

150

200

250

300

350

none supp (5%) 10% 20% 40%

Candidates Rules (before Pruning) RulesTime (Rule Base) Time (Mining)

247

184 179167

123

44 42 40 40 3922 22 20 20 19

7.25

5.985.73

5.21

3.89

5.82

4.013.89 3.81

2.95

0

1

2

3

4

5

6

7

8

0

50

100

150

200

250

300

none supp (5%) 10% 20% 40%

Candidates Rules (before Pruning) RulesTime (Rule Base) Time (Mining)

NUMBER OF RULES TIME (SEC) NUMBER OF RULES TIME (SEC)

(a) Results using rule template set 1

300

243241 229

222

22 22 22 20 1914 14 14 12 11

7.75

6.44 6.446.12

5.36.74

5.445.23 5.08 4.9

0

1

2

3

4

5

6

7

8

9

0

50

100

150

200

250

300

350

none supp (5%) 10% 20% 40%

Candidates Rules (before Pruning) RulesTime (Rule Base) Time (Mining)

247

184 179167

123

44 42 40 40 3922 22 20 20 19

7.25

5.985.73

5.21

3.89

5.82

4.013.89 3.81

2.95

0

1

2

3

4

5

6

7

8

0

50

100

150

200

250

300

none supp (5%) 10% 20% 40%

Candidates Rules (before Pruning) RulesTime (Rule Base) Time (Mining)

NUMBER OF RULES TIME (SEC) NUMBER OF RULES TIME (SEC)

(b) Results using rule template set 2

Figure 5: Performance evaluation using different sets of rule templates

and the approval of university business trips as well as the management of

accommodations and transfers, e.g., booking hotels and transport tickets. The

system has been used for 6 months by 11 employees of a research institute520

of the University of Bayreuth (Germany). The organisational model of the

institute assigns the 11 identities to 4 distinct roles, specifically 6 PhD students,

1 professor, 1 secretary and 3 administration employees. In total, there are 128

business trips, i.e., traces, recorded. All the computation times reported in this

section are measured on a Core i7 CPU @2.80 GHz with 8 GB Ram.525

Our approach has been tested with two different sets of rule templates.

Fig. 5a shows the results of applying the approach with template set 1, which

contains the templates direct, role, binding and orgDistMulti. Fig. 5b shows

the results for template set 2, which contains the sequence template and the

roleSequence cross-perspective template.530

We analysed the time to build the rete network, i.e., the rule base7, as

well as the time to perform the actual mining process taking into account a

different number of rule candidates. This was achieved by considering different

minSupp values during the pre-processing phase ranging from 0 to 0.4 (cf.

7Note that the rule base only needs to be built once for different applications since the set

of candidates depends on the occurring entities and not on the number of events or traces.

23

Section 4). The analysis shows the feasibility of our approach since in both535

tests, despite a big amount of candidates, only a manageable number of rules has

been discovered. Especially the diagram in Fig. 5b highlights the benefit of the

pre-processing approach. With increasing minSupp, the number of candidates

to check considerably decreases, which reduces the processing time up to 50%.

However, almost the same number of rules has been discovered in all cases before540

the post-processing phase. However, both diagrams show that the number of

extracted rules is clearly reduced by pruning unnecessary rules. Fig. 5b, e.g.,

shows that the number of rules can be reduced by 50%.

In order to check the efficiency of our approach we also applied the imple-

mentation of the DeclareMiner [30] available in the Process Mining Framework545

(ProM) by only analysing the precedence template of Declare [18], which equates

to the sequence template of DPIL. With standard settings, the DeclareMiner

needed 14.85 sec to analyse the provided event log with the precedence template.

Even if we analysed the example log with 2, respectively 4, rule templates, our

approach was still faster in any case. For template set 1 and without pre-550

processing, the generation of the rule base for the rete algorithm took 7.75 sec

while the actual analysis took only 6.74 sec.

6.3. Application to Real-Life Event Log

In this section we describe our findings when applying the approach to the

university business trip log of Section 6.2. We analysed the log with the 6 afore-555

mentioned rule templates. With minSupp = 0.1 in the pre-processing phase and

after removing unnecessary rules in the post-processing phase, we extracted 34

rules in total. The extracted resource assignment rules are composed of 4 direct,

1 role, 5 binding and 4 orgDistMulti rules. The rules with control-flow infor-

mation include 14 sequence and 6 roleSequence rules. For the classification in560

satisfied and violated rules, we used minConf = 0.85 and minInt = 1.0. For

space reasons, we only describe some interesting parts of the resulting model

(cf. Figure 6). The discovered model shows that task “Approve Application”

has mostly been performed by the identity “SJ” (direct). Furthermore, “Check

24

ensure direct(Approve Application, SJ)

ensure role(Check Application, Administration)

ensure binding(Apply for trip, Book flight)

ensure binding(Apply for trip, Book accommodation)

ensure binding(Apply for trip, Book transfer)

ensure orgDistMulti(Approve Application, Apply for trip, supervisor)

ensure roleSequence(Apply for trip, Book flight, Student)

Figure 6: Examples of discovered rules

Application” has mostly been performed by a resource with the role “Admin-565

istration” (role). The three binding of duties rules show that the resource who

booked the flight tickets, the accommodation and the transfer service has to

be the person that applies for the trip (binding). Moreover, the resource who

approves the trip application is the supervisor of the applicant (orgDistMulti).

Regarding cross-perspective patterns, there are cases in which certain employees570

already booked a flight without applying for the trip. However, when analysing

the task order under consideration of performing resources, we extracted that

students always applied for the trip before they booked the flight (roleSequence).

In a second step we evaluated the quality of the mining results and how vary-

ing the mining configuration, i.e., different thresholds, influences it. Therefore,575

three discovered models (M1, M2, M3) based on different configurations of the

approach on the same event log were discussed and evaluated in a workshop. The

models were extracted using different minSupp values during the pre-processing

phase as well as different minConf values during the mining phase. Table 4

shows the characteristics of the discovered models. M1 has been discovered by580

applying the approach without any pre-processing (low filtering). M2 depicts

the model that has been described before and is based on a pre-processed log

with minSupp=0.1 (medium filtering). Both M1 and M2 include rules r with

conf(r) > 0.85. One task that occurs in less than 10% of traces and the corre-

sponding rules have been filtered in M2. M3 is based on minConf = 0.9, i.e.,585

less rules are classified as satisfied (high filtering). The workshop was carried

out with 8 process participants, i.e., university employees that represented all

25

the organisational groups involved. After we provided a general overview about

the process and the workshop setting, each of the extracted rules was classified

by the participants.590

For evaluating the quality of the results, we rely on standard metrics from

information retrieval precision and recall [31]. The harmonic mean (F-measure)

of precision and recall is an adequate value for measuring the overall quality of

extracted models [31]. To compute recall and precision, rules have been classified

into one of three categories, i.e., (i) true-positive (TP : correctly discovered); (ii)

false-positive (FP : incorrectly discovered); (iii) false-negative (FN : incorrectly

missing). Precision, recall and F-measure are defined as follows:

Precision =TP

TP + FP, Recall =

TPTP + FN

, F = 2 · P ·RP +R

(4)

The results of the workshop as well as the calculated quality metrics are

collected in Table 4. First of all, we focus on the results of M2. According to

Model 1 Model 2 Model 3

Mining configuration Low filtering Medium filtering High filtering

minSupp (Pre-processing) � 0.1 0.1

minSupp � 0.2 0.2

minConf 0.85 0.85 0.9

Characteristics of models

Number of tasks 10 9 9

Number of identities 10 10 10

Number of rules 47 39 31

Metrics

TP (correctly discovered) 40 34 28

FP (incorrectly discovered) 7 5 3

FN (incorrectly missing) 0 6 12

Precision 0.85 0.87 0.9

Recall 1.0 0.85 0.7

F-measure 0.92 0.86 0.8

Table 4: Characteristics, results and metrics of discovered models

26

the information gathered from the process participants, 34 of 39 rules have been

classified as relevant (TP ) while 5 rules have been discovered incorrectly (FP ).

Furthermore, 6 missing rules (FN ) have been identified in the discussion. The595

reason is that a task and the assignment rules related to that task were filtered in

the pre-processing phase. Based on this classification, we obtain Precision=0.87,

Recall=0.85 and therefore, F=0.86. Comparing the three models in Table 4 we

can observe that M1 has the highest F-measure, i.e., the best quality. Since the

model was extracted without filtering infrequent behaviour, there were no rules600

missing (Recall=1.0 ). Without filtering, however, M1 also contains 7 irrelevant

rules leading to a lower precision value. Since M3 is based on a higher minConf

threshold, the model contains fewer rules. However, some of the missing rules

were identified as relevant by the workshop participants. Due to the missing

rules, M3 features the lowest recall and thus also the lowest F-measure.605

7. Related Work

Several approaches have been proposed in the literature for the discovery of

declarative process models. In [25] the authors present an approach that allows

the user to select from a set of predefined Declare templates the ones to be used

for the discovery. Maggi et al. propose an evolution of this approach in [24] to610

improve performance by pre-processing the event log with frequent pattern min-

ing techniques. Other approaches to improve the performance of process mining

are presented in [3, 32]. Additionally, there are post-processing approaches that

aim at simplifying the resulting Declare models in terms of redundancy elimi-

nation [33] and disambiguation [27]. The approach proposed in [34] allows for615

the specification of rules that go beyond the traditional Declare templates. In

[35], an approach for analysing event logs with Timed Declare, an extension of

Declare that relies on timed automata, is described. The work in [36] first cov-

ered the data perspective in declarative process mining, although this approach

only allows for the discovery of discriminative activation conditions. In essence,620

the focus of the aforementioned approaches is control-flow with extensions to

27

cover data without analysing resource-related information.

Complementary to them are techniques for mining the organisational per-

spective of a process [33]. Methods for analysing event logs w.r.t. resources

are mainly focused on enriching a given procedural model with resource assign-625

ments [13]. Several methods focus on extracting an organisational model [9]

or a social network [8]. There are also approaches that analyse the influence

of resources on process performance [10]. However, the approaches that are

of highest interest to us are those collected in Table 5, which address the dis-

covery of organisational or cross-perspective patterns. Staff assignment mining630

[37] is able to extract complex assignment rules based on decision tree learn-

ing. However, the resource assignments are only related to one single task (cf.

Section 5.1.1). Works on role mining [11, 12] are, on the contrary, interested

in those types of rules referring to several tasks (cf. Section 5.1.2) but disre-

gard other patterns. Resource mining is also implemented in ProM. In [38] the635

authors propose a two-step technique for enriching a given control-flow model

with swimlanes based on the Handover of Roles (HooR) principle8. In the first

step the pairs of immediately consecutive activities are analyzed in terms of po-

tential role changes based on three rules: (i) pairs of immediately consecutive

activities that are always executed by the same resource do not involve a HooR,640

(ii) pairs of immediately consecutive activities that are each executed by exactly

the same set of resources do not involve a HooR and (iii) pairs of immediately

consecutive activities that are, to a certain proportion w, executed by the same

resources do not involve a HooR. All rules are based upon the assumption that

each resource has exactly one role. The clustering-inspired algorithm generates645

a partition of activities for each HooR. The last step of the algorithm merges

similar partitions in order to identify the actual roles. Finally, the algorithm

chooses the most suitable final partitioning based on an entropy measure.

None of the aforementioned approaches on resource mining covers the whole

sets of organisational and cross-perspective patterns that constitute the goal650

8For space limitations, we refer to [38] for details on this principle.

28

of our work. The DpilMiner was developed to bridge that gap and hence, we

used its mining approach [16] for the mining phase of our framework, which

we extended with pre-processing and post-processing techniques inspired by the

solutions related to mining the process control-flow.

8. Conclusions and Future Work655

In this paper we presented a process mining framework to discover resource-

aware process models. Our approach is based upon the mining approach in-

troduced in [16], which we extended with pre-processing and post-processing

phases. This increased efficiency while generating simplified process models that

provide the same valuable information, as demonstrated by our evaluations.660

Since our approach relies on DPIL [20], the mining capabilities are limited to

its expressiveness. Therefore, inter-case dependencies, such as those represented

in the History-Based Distribution pattern, cannot be discovered. It is an inter-

esting question for future research how such dependencies can be mined and

effectively depicted in a process model. Furthermore, there might be more ways665

to prune discovered models that take into account more knowledge besides hier-

archies and transitive reduction. By pruning more intelligently, a better model

Pattern Mining approach

Direct Distribution [9, 12, 16, 37, 39, 38]

Role-based Distribution [9, 12, 16, 37, 39, 38]

Deferred Distribution -

Separation of Duties [11, 12, 16]

Case Handling [9, 12, 16]

Retain Familiar [11, 12, 16]

Capability-based Distribution [37, 16]

History-based Distribution -

Organisational Distribution [37] (single task) [16] (incl. several tasks)

Cross-Perspective Patterns [16]

Table 5: Existing approaches for mining the organisational perspective

29

could be obtained. Finally, we plan to investigate options for mapping the

output to graphical process modelling notations to increase readability.

References670

[1] M. Dumas, M. L. Rosa, J. Mendling, H. A. Reijers, Fundamentals of

Business Process Management, Springer-Verlag Berlin Heidelberg, 2013.

doi:10.1007/978-3-642-33143-5.

[2] W. van der Aalst, Process mining: discovery, conformance and enhance-

ment of business processes, Springer-Verlag Berlin Heidelberg, 2011. doi:675

10.1007/978-3-642-19345-3.

[3] C. Di Ciccio, M. Mecella, On the Discovery of Declarative Control Flows

for Artful Processes, ACM Trans. Management Inf. Syst. 5 (4) (2015) 24:1–

24:37. doi:10.1145/2629447.

[4] W. M. P. van der Aalst, M. Rosemann, M. Dumas, Deadline-based es-680

calation in process-aware information systems, Decision Support Systems

43 (2) (2007) 492–511. doi:10.1016/j.dss.2006.11.005.

[5] W. M. P. van der Aalst, K. M. van Hee, J. M. E. M. van der Werf, A. Ku-

mar, M. Verdonk, Conceptual model for online auditing, Decision Support

Systems 50 (3) (2011) 636–647. doi:10.1016/j.dss.2010.08.014.685

[6] M. de Leoni, M. Adams, W. M. P. van der Aalst, A. H. M. ter Hofstede,

Visual support for work assignment in process-aware information systems:

Framework formalisation and implementation, Decision Support Systems

54 (1) (2012) 345–361. doi:10.1016/j.dss.2012.05.042.

[7] C. Cabanillas, D. Knuplesch, M. Resinas, M. Reichert, J. Mendling,690

A. Ruiz-Cortes, RALph: A Graphical Notation for Resource Assign-

ments in Business Processes, in: Int. Conf. on Advanced Information Sys-

tems Engineering (CAiSE), Vol. 9097, 2015, pp. 53–68. doi:10.1007/

978-3-319-19069-3_4.

30

[8] W. van der Aalst, H. A. Reijers, M. Song, Discovering Social Networks695

from Event Logs, Computer Supported Cooperative Work 14 (6) (2005)

549–593. doi:10.1007/s10606-005-9005-9.

[9] M. Song, W. van der Aalst, Towards comprehensive support for organi-

zational mining, Decision Support Systems 46 (1) (2008) 300–317. doi:

10.1016/j.dss.2008.07.002.700

[10] J. Nakatumba, W. van der Aalst, Analyzing resource behavior using process

mining, in: Business Process Management Workshops, 2010, pp. 69–80.

doi:10.1007/978-3-642-12186-9_8.

[11] M. Leitner, A. Baumgrass, S. Schefer-Wenzl, S. Rinderle-Ma, M. Strem-

beck, A Case Study on the Suitability of Process Mining to Produce705

Current-State RBAC Models, in: Business Process Management Work-

shops, 2012, pp. 719–724. doi:10.1007/978-3-642-36285-9_72.

[12] A. Baumgrass, M. Strembeck, Bridging the gap between role mining and

role engineering via migration guides, Inf. Sec. Techn. Report 17 (4) (2013)

148–172. doi:10.1016/j.istr.2013.03.003.710

[13] W. Zhao, X. Zhao, Process Mining from the Organizational Perspective,

in: Advances in Intelligent Systems and Computing, Vol. 277, 2014, pp.

701–708. doi:10.1007/978-3-642-54924-3_66.

[14] N. Russell, W. M. P. van der Aalst, A. H. M. ter Hofstede, D. Edmond,

Workflow Resource Patterns: Identification, Representation and Tool Sup-715

port, in: Advanced Information Systems Engineering, 2005, pp. 216–232.

doi:10.1007/11431855_16.

[15] M. de Leoni, W. M. van der Aalst, M. Dees, A general process min-

ing framework for correlating, predicting and clustering dynamic behav-

ior based on event logs, Information Systems 56 (2016) 235–257. doi:720

10.1016/j.is.2015.07.003.

31

[16] S. Schonig, C. Cabanillas, S. Jablonski, J. Mendling, Mining the Organisa-

tional Perspective in Agile Business Processes, in: Int. Conf. on Enterprise,

Business-Process and Information Systems Modeling (BPMDS), Vol. 214 of

LNBIP, Springer, 2015, pp. 37–52. doi:10.1007/978-3-319-19237-6_3.725

[17] E. Verbeek, J. Buijs, B. van Dongen, W. van der Aalst, XES, xESame,

and ProM 6, in: Information Systems Evolution, Vol. 72, 2011, pp. 60–75.

doi:10.1007/978-3-642-17722-4_5.

[18] W. van der Aalst, M. Pesic, H. Schonenberg, Declarative workflows: Bal-

ancing between flexibility and support, Computer Science - R&D 23 (2)730

(2009) 99–113. doi:10.1007/s00450-009-0057-9.

[19] OMG, BPMN 2.0, Recommendation, OMG (2011).

[20] M. Zeising, S. Schonig, S. Jablonski, Towards a Common Platform for the

Support of Routine and Agile Business Processes, in: IEEE Int. Conf.

on Collaborative Computing: Networking, Applications and Worksharing,735

2014, pp. 94–103. doi:10.4108/icst.collaboratecom.2014.257269.

[21] M. Montali, Specification and Verification of Declarative open Interaction

Models - A logic-based approach, Vol. 56, Springer, 2010. doi:10.1007/

978-3-642-14538-4.

[22] F. Maggi, M. Montali, M. Westergaard, W. van der Aalst, Monitor-740

ing Business Constraints with Linear Temporal Logic: An Approach

Based on Colored Automata, in: Int. Conf. on Business Process Man-

agement (BPM), Vol. 6896, Springer, 2011, pp. 132–147. doi:10.1007/

978-3-642-23059-2_13.

[23] C. Bussler, Organisationsverwaltung in Workflow-Management-Systemen,745

Deutscher Universitatsverlag, 1998. doi:10.1007/978-3-663-08832-5.

[24] F. M. Maggi, J. C. Bose, W. van der Aalst, Efficient Discovery of Under-

standable Declarative Process Models from Event Logs, in: Int. Conf. on

32

Advanced Information Systems Engineering (CAiSE), Vol. 7328, 2012, pp.

270–285. doi:10.1007/978-3-642-31095-9_18.750

[25] F. M. Maggi, A. Mooij, W. van der Aalst, User-Guided Discovery of Declar-

ative Process Models, in: IEEE Symposium on Computational Intelligence

and Data Mining, 2011, pp. 192–199. doi:10.1109/CIDM.2011.5949297.

[26] R. J. C. Bose, W. M. van der Aalst, Analysis of Patient Treatment Pro-

cedures, in: Business Process Management Workshops, Vol. 99, 2011, pp.755

165–166. doi:10.1007/978-3-642-28108-2_17.

[27] F. M. Maggi, J. C. Bose, W. van der Aalst, A Knowledge-Based Integrated

Approach for Discovering and Repairing Declare Maps, in: Int. Conf. on

Advanced Information Systems Engineering (CAiSE), Vol. 7908, 2013, pp.

433–448. doi:10.1007/978-3-642-38709-8_28.760

[28] A. V. Aho, M. R. Garey, J. D. Ullman, The Transitive Reduction of a

Directed Graph, SIAM J. Comput. 1 (2) (1972) 131–137. doi:10.1137/

0201008.

[29] C. Forgy, Rete: A Fast Algorithm for the Many Patterns/Many Ob-

jects Match Problem, Artif. Intell. 19 (1) (1982) 17–37. doi:10.1016/765

0004-3702(82)90020-0.

[30] F. M. Maggi, Declarative Process Mining with the Declare Component

of ProM, in: Business Process Management Demos, Vol. 1021 of CEUR

Workshop Proceedings, 2013.

URL http://ceur-ws.org/Vol-1021/paper_8.pdf770

[31] A. Rozinat, A. K. A. de Medeiros, C. W. Gunther, A. Weijters, W. M.

van der Aalst, The Need for a Process Mining Evaluation Framework in

Research and Practice, in: Business Process Management Workshops, Vol.

4928, 2008, pp. 84–89. doi:10.1007/978-3-540-78238-4_10.

[32] M. Westergaard, C. Stahl, H. Reijers, UnconstrainedMiner: Efficient Dis-775

covery of Generalized Declarative Process Models, Tech. Rep. 13-28, Eind-

33

hoven University of Technology (2013).

URL https://publications.hse.ru/en/preprints/117624631

[33] J. C. Bose, F. M. Maggi, W. van der Aalst, Enhancing Declare Maps Based

on Event Correlations, in: Int. Conf. on Business Process Management780

(BPM), Vol. 8094, 2013, pp. 97–112. doi:10.1007/978-3-642-40176-3_

9.

[34] F. Chesani, E. Lamma, P. Mello, M. Montali, F. Riguzzi, S. Storari, Ex-

ploiting inductive logic programming techniques for declarative process

mining, Trans. Petri Nets and Other Models of Concurrency 2 (2009) 278–785

295. doi:10.1007/978-3-642-00899-3_16.

[35] F. M. Maggi, Discovering Metric Temporal Business Constraints from

Event Logs, in: Int. Conf. on Perspectives in Business Informatics Re-

search (BIR), Vol. 194, Springer, 2014, pp. 261–275. doi:10.1007/

978-3-319-11370-8_19.790

[36] F. M. Maggi, M. Dumas, Discovering Data-Aware Declarative Process

Models from Event Logs, in: Int. Conf. on Business Process Management

(BPM), Vol. 8094, 2013, pp. 1–16. doi:10.1007/978-3-642-40176-3_8.

[37] S. Rinderle-Ma, W. M. van der Aalst, Life-cycle support for staff assign-

ment rules in process-aware information systems, Tech. rep., Eindhoven795

University of Technology (2007).

URL http://dbis.eprints.uni-ulm.de/373/

[38] A. Burattin, A. Sperduti, M. Veluscek, Business models enhancement

through discovery of roles, in: IEEE Symposium on Computational In-

telligence and Data Mining, 2013, pp. 103–110. doi:10.1109/CIDM.2013.800

6597224.

[39] T. Jin, J. Wang, L. Wen, Organizational modeling from event logs, in:

Int. Conf. on Grid and Cooperative Computing (GCC), 2007, pp. 670–675.

doi:10.1109/GCC.2007.93.

34